Multi-Scale Graph Convolutional Interaction Network For Salient Object Detection

Remarkable progress has been achieved for salient object detection based on deep learning. However, most of the previous works have the issues of how to extract more effective information from scale-varying data and how to improve the boundary quality. In this paper, we propose the multi-scale graph convolutional interaction network (MGCINet), which consists of the feature interaction module (FIM), the feature aggregation module (FAM), and the residual refinement module (RRM). FIMs fuse interactive features from neighboring scales. Based on two-layers graph convolutional network, FAMs aggregate scale-specific information by graph nodes interaction. RRMs optimize the coarse saliency maps with blurred boundaries by U-net residual blocks. In addition, we propose multi-scale weighted structural loss to assign different weights to pixels while focusing on image structure at various scales. Experiments show that our method outperforms the state-of-the-arts on five benchmark datasets under different evaluation metrics.